[jason-fxz]

Muse-Glimmer fused its per-head attention output gate into the q/k/v projection (qkvg_proj). The gate is now its own linear next to qkv_proj, as vLLM builds it (qkv_proj + output_gate_proj), in every format the family loads (bf16, compressed-tensors NVFP4, GGUF).

The reason is the NVFP4 release: Muse-Glimmer-30B-NVFP4 quantizes the gate with a different weight global scale than q/k/v in all 52 layers (the gate's global is down to 0.40x of q/k/v's), while q/k/v share one. main's Triton W4A16 linear keeps one global per output row, so the fused weight is exact there, but GEMMs that take one global per weight (cuBLASLt scaled_mm, FlashInfer mm_fp4, used by #737) cannot run a fused weight whose parts differ without scaling one part wrongly. Across the local NVFP4 checkpoints (Qwen3.6-27B, Qwen3.8-27B, Gemma-4-31B, Muse-Glimmer-30B, GLM-4.7) this was the only fused weight with differing per-part globals. Merge before #737.

Changes

• models/muse_glimmer/attention.py: qkv_proj (q, k, v) and gate_proj (the gate) instead of qkvg_proj.
• Loaders: the bf16 and compressed-tensors NVFP4 readers fuse only q/k/v (.self_attn.gate_proj loads on its own; the full suffix keeps it out of the MLP's gate/up fusion), the GGUF reader writes qkv_proj and gate_proj; the family uses the shared qkv_proj / gate_up_proj packed mapping.
• Tests and docs/models.md.

Breaking: an FTW checkpoint converted with qkvg_proj no longer loads; reconvert it with ft checkpoint.

Results

On RTX PRO 6000, main's W4A16 linears:

• Muse-Glimmer-30B-NVFP4, greedy 96 tokens on a three-sentence prompt: the same text as main.
• bs=1 decode, 256 greedy tokens, two server runs x 3 requests each: 72.0 -> 71.0 tok/s (-1.4%), from one more GEMM per layer.
• GGUF: synthetic 2-layer muse-glimmer GGUFs (q / k Q8_0 + v Q4_0 + gate F16; all Q8_0; a fused attn_qkv + Q4_0 gate) give qkv_proj and gate_proj outputs equal to the old qkvg_proj output bitwise at T = 1 / 7 / 64 (an F16 gate now runs as a plain bf16 linear, the same bf16 values). There is no real Muse-Glimmer GGUF on this box.

Tests

pytest tests/models/test_muse_glimmer.py tests/models/test_muse_glimmer_vision.py tests/server/test_muse_glimmer_parsers.py tests/models/test_quant_config.py tests/models/test_models_registry.py tests/checkpoint: 235 passed, 12 skipped, 5 failed; the 5 failures (test_quant_config.py, from the local HF cache) also fail on main. The test_quant_config.py Muse-Glimmer cases that build the model on the bf16, FP8 block and NVFP4 checkpoints on disk pass with the separate gate.